US 8176493 B1
Management of contexts that execute on a computer system is described. More specifically, context scheduling in a virtual machine environment is described. A set of coscheduled contexts is monitored. If a skew metric associated with a first context of the coscheduled contexts fails to satisfy a condition, then a subset of the coscheduled contexts is descheduled although the first context remains scheduled.
1. A method of managing contexts that execute on a computer system, said method comprising:
monitoring a plurality of contexts in a coscheduling grouping for unsynchronized scheduling of said plurality of contexts;
determining that a skew metric associated with a first context of said plurality of contexts fails to satisfy a condition, said skew metric quantifying skew in unsynchronized scheduling for the first context of said plurality of contexts;
determining if a relaxed costop should be performed based on said skew metric associated with said first context and a skew metric associated with a subset of contexts in said plurality of contexts; and
if said relaxed costop is determined, descheduling said subset of said plurality of contexts while said first context remains scheduled, wherein said subset of said plurality of contexts comprises all of said plurality of contexts except for said first context.
2. The method of
3. The method of
4. The method of
5. The method of
6. The method of
7. The method of
8. The method of
9. The method of
10. The method of
11. The method of
12. The method of
13. The method of
determining if a strict costop or said relaxed costop should be performed based on said skew metric for said first context and said skew metric for said subset of contexts; and
descheduling said subset of contexts and said first context if said strict costop is determined.
14. The method of
determining a type of instruction being executed by said plurality of contexts; and
determining if said strict costop or said relaxed costop should be performed based on said type of instruction.
15. A non-transitory computer-readable medium having computer-executable instructions for causing a computer system to perform a method of managing contexts, said method comprising:
monitoring a plurality of contexts in a coscheduling grouping for unsynchronized scheduling of said plurality of contexts;
identifying a first context that has associated therewith a skew metric that fails to satisfy a condition, said skew metric quantifying skew in unsynchronized scheduling for first context of said plurality of contexts;
determining if a relaxed costop should be performed based on said skew metric associated with said first context and a skew metric associated with a subset of contexts in said plurality of contexts; and
if said relaxed costop is determined, descheduling a said subset of contexts while said first context remains scheduled, wherein said subset of said plurality of contexts comprises all of said plurality of contexts except for said first context.
16. The non-transitory computer-readable medium of
17. The non-transitory computer-readable medium of
18. A computer system comprising:
a hardware platform comprising a processor and a memory, said hardware platform having installed thereon a plurality of virtual machines each comprising a plurality of virtual processors, said hardware platform also having installed thereon virtualization software that serves as an interface between said processor and memory and said virtual machines;
said virtualization software comprising instructions operable to cause said processor to be operable to:
monitor a plurality of contexts in a coscheduling grouping for unsynchronized scheduling of said plurality of contexts;
determine that a skew metric associated with a first context of said plurality of contexts fails to satisfy a condition, said skew metric quantifying skew in unsynchronized scheduling for first context of said plurality of contexts;
determine if a relaxed costop should be performed based on said skew metric associated with said first context and a skew metric associated with a subset of contexts in said plurality of contexts; and
if said relaxed costop is determined, deschedule said subset of said plurality of contexts while said first context remains scheduled, wherein said subset of said plurality of contexts comprises all of said plurality of contexts except for said first context.
19. The computer system of
20. The computer system of
21. The computer system of
22. The computer system of
23. The computer system of
24. The computer system of
25. The computer system of
26. The computer system of
1. Field of the Invention
This invention relates to context scheduling in general, and more particularly to processor scheduling in a virtual machine environment.
2. Description of the Related Art
As is well known in the field of computer science, a virtual machine (VM) is an abstraction—a “virtualization”—of an actual physical computer system. The advantages of virtual machine technology have become widely recognized. Among these advantages is the ability to run multiple virtual machines on a single host platform. This makes better use of the capacity of the hardware, while still ensuring that each user enjoys the features of a “complete” computer. Depending on how it is implemented, virtualization can also provide greater security, since the virtualization can isolate potentially unstable or unsafe software so that it cannot adversely affect the hardware state or system files required for running the physical (as opposed to virtual) hardware.
A virtual machine or “guest” is installed on a “host platform,” or simply “host,” which includes system hardware and one or more layers or co-resident components comprising system-level software, such as an operating system or similar kernel, or a virtual machine monitor or hypervisor, or some combination of these. Each VM will typically have both virtual system hardware and guest system software. A single VM may be configured with more than one virtualized processor. To permit computer systems to scale to larger numbers of concurrent threads, systems with multiple processors have been developed. These symmetric multi-processor (SMP) systems are available as extensions of the personal computer platform. Essentially, an SMP system is a hardware platform that connects multiple processors to a shared main memory and shared input/output devices. Virtual machines may also be configured as SMP VMs.
In the field of processor scheduling, the ability to simultaneously schedule multiple “contexts” (e.g., virtual processors) is commonly referred to as “coscheduling” or “gang scheduling.” There are a number of reasons why coscheduling is beneficial. For instance, coscheduling is generally used to ensure that when multiple, cooperating contexts need to communicate or synchronize with each other, then they will all be ready to do so at the same time. When the contexts are virtual processors associated with a virtual machine, coscheduling is also desirable to maintain the illusion presented to the guest operating system that it is running on a dedicated physical multiprocessor. Synchronous execution may improve performance significantly, and may even be required for correctness in some cases where inter-processor operations are expected to complete quickly.
However, strict coscheduling can cause fragmentation, with the result that some physical processors may remain idle or may be underutilized even when the overall demand for processor time is high. For example, consider the problem of scheduling two CPU-bound virtual machines on a physical machine with two processors: VM1, a uniprocessor VM with a single virtual processor, and VM2, an SMP VM with two virtual processors that must be coscheduled. When VM1 is scheduled, VM2 cannot run because both of its virtual processors must be coscheduled but only one physical processor is available. As a result, even though the second physical processor is available, it will remain idle, even though there is a VM ready to run (e.g., VM2). Thus, the computer system's resources are not efficiently utilized, reducing overall performance.
Management of contexts that execute on a computer system is described. More specifically, context scheduling in a virtual machine environment is described. Times at which a context transitions from a scheduled state to a descheduled state and times at which the context transitions from a descheduled state to a scheduled state are recorded for each context. Skew is detected using the recorded times. The amount of skew can be quantified, and a corrective action is triggered if the amount of skew fails to satisfy a threshold value.
The present invention pertains generally to the tracking and management of coscheduled “contexts” that execute on a computer system. A context may represent different software constructs, depending on the underlying system. A context may be a process associated with an application, or a thread that is associated with a process.
A context may instead be a virtual machine or a “virtual processor” associated with a virtual machine. In one embodiment, the present invention pertains to coscheduling a group of virtual machines. For example, two virtual machines may be cooperating servers as part of a multi-tier service that would be improved if both of the virtual machines are coscheduled (regardless of the number of virtual processors per virtual machine). Similarly, the present invention pertains to coscheduling any group of applications that collaborate on the host (such as a database server and an application server that work together to process a transaction). The context sets may be permanent (e.g., as in the case of virtual processors that are part of a virtual machine) or transient (e.g., as in the case of virtual machines or applications that collaborate on a transaction only until the transaction is completed).
“Coscheduling” and “gang scheduling” refer to the concurrent or simultaneous scheduling of a set of cooperating contexts that may need to communicate and/or synchronize with one another. As will be seen, the present invention allows “relaxed” coscheduling in addition to “strict” coscheduling, and in particular the application of relaxed coscheduling to non-batch tasks/contexts. Thus, although coscheduled contexts are intended to be scheduled (e.g., executed) simultaneously (strict coscheduling), according to the invention some contexts within the set of coscheduled contexts may be scheduled while others are temporarily descheduled (relaxed coscheduling). Accordingly, depending on the manner in which the term “coscheduled contexts” is used, it may herein refer either to cooperating contexts that are scheduled concurrently, or to cooperating contexts that are not scheduled concurrently because of a policy of relaxed coscheduling. Coscheduled or cooperating contexts may also be referred to “sibling contexts.”
If there is a mix of scheduled and descheduled contexts, so that some contexts in a set of coscheduled contexts are not executing at the same time as others in the set, then their execution is said to be “skewed.” The notion of coscheduling “skew” is intended to quantify the amount of unsynchronized scheduling or execution between coscheduled contexts.
The present invention pertains more specifically to the definition, detection, measurement and management of skew between coscheduled contexts. The present invention introduces several mechanisms and policies that track the amount of skew during periods of relaxed coscheduling and that, when warranted, implement corrective actions to reduce the amount of skew. For example, the amount of skew per context may be defined as the amount of time that a context was descheduled while other contexts in the set of coscheduled contexts were scheduled. If a condition applied to the amount of skew is not satisfied (e.g., a predefined threshold value is exceeded), then one or more corrective actions can be implemented. Corrective actions introduced herein include skew reduction policies referred to as “costop,” “costart” and “coswap.” A costop policy may include descheduling some contexts (relaxed costop) or all contexts (strict costop) in the set of coscheduled contexts. A costart policy may include scheduling some (relaxed costart) or all (strict costart) contexts in the set of coscheduled contexts to begin execution at the same time. A coswap policy may include rapidly and alternately scheduling and descheduling contexts on a physical processor, so that each context makes forward progress and does not get too far ahead or behind its siblings, thereby bounding the amount of skew. Additional information regarding the definition, detection, measurement and management of skew between coscheduled contexts is provided in conjunction with
Each VM 200 will typically have both virtual system hardware 201 and guest system software 202. The virtual system hardware typically includes at least one virtual central processing unit (CPU) 210, virtual memory 230, at least one virtual disk 240, and one or more virtual devices 270. Note that a disk—virtual or physical—is also a “device,” but is usually considered separately because of the important role of the disk. All of the virtual hardware components of the VM 200 may be implemented in software using known techniques to emulate the corresponding physical components. The guest system software 202 includes a guest operating system (OS) 220 and drivers (DRVS) 224 as needed for the various virtual devices 270.
A single VM 200 may be configured with more than one virtualized processor. To permit computer systems to scale to larger numbers of concurrent threads, systems with multiple CPUs have been developed. These symmetric multi-processor (SMP) systems are available as extensions of the PC platform and from other vendors. Essentially, an SMP system is a hardware platform that connects multiple processors to a shared main memory and shared I/O devices. Virtual machines may also be configured as SMP VMs.
Yet another configuration is found in a so-called “multi-core” architecture, in which more than one physical CPU is fabricated on a single chip, with its own set of functional units (such as a floating-point unit and an arithmetic/logic unit ALU), and can execute threads independently; multi-core processors typically share only very limited resources, such as some cache. Still another technique that provides for simultaneous execution of multiple threads is referred to as “simultaneous multi-threading” (SMT), in which more than one logical CPU (hardware thread) operates simultaneously on a single chip, but in which the logical CPUs flexibly share some resource such as caches, buffers, functional units, etc. This invention may be used regardless of the type—physical and/or logical—or number of processors included in a VM.
If the VM 200 is properly designed, applications 260 running on the VM will function as they would if run on a “real” computer, even though the applications are running at least partially indirectly, that is via the guest OS 220 and virtual processor(s) 210-0, 210-1, . . . , 210-m. Executable files will be accessed by the guest OS 220 from the virtual disk 240 or virtual memory 230, which will be portions of the actual physical disk 140 or memory 130 allocated to that VM. Once an application 260 is installed within the VM 200, the guest OS 220 retrieves files from the virtual disk 240 just as if the files had been pre-stored as the result of a conventional installation of the application. The design and operation of virtual machines are well known in the field of computer science.
Some interface is generally required between the guest software within a VM 200 and the various hardware components and devices in the underlying hardware platform 100. This interface—which may be referred to generally as “virtualization software”—may include one or more software components and/or layers, possibly including one or more of the software components known in the field of virtual machine technology as “virtual machine monitors” (VMMs) 300, . . . , 300-n, “hypervisors,” or virtualization “kernels” 600. Because virtualization terminology has evolved over time and has not yet become fully standardized, these terms do not always provide clear distinctions between the software layers and components to which they refer. For example, “hypervisor” is often used to describe both a VMM 300 and a kernel 600 together, either as separate but cooperating components or with one or more VMMs incorporated wholly or partially into the kernel itself; however, “hypervisor” is sometimes used instead to mean some variant of a VMM 300 alone, which interfaces with some other software layer(s) or component(s) to support the virtualization. Moreover, in some systems, some virtualization code is included in at least one “superior” VM (e.g., VM 200) to facilitate the operations of other VMs. Furthermore, specific software support for VMs 200, . . . , 200-n may be included in the host OS itself. Unless otherwise indicated, the invention described below may be used in virtualized computer systems having any type or configuration of virtualization software.
The various virtualized hardware components in the VM, such as the virtual CPU(s) 210-0, . . . , 210-m, the virtual memory 230, the virtual disk 240, and the virtual device(s) 270, are shown as being part of the VM 200 for the sake of conceptual simplicity. In actuality, these “components” are usually implemented as software emulations 370 included in the VMM 300. One advantage of such an arrangement is that the VMM may (but need not) be set up to expose “generic” devices, which facilitate VM migration and hardware platform-independence.
Different systems may implement virtualization to different degrees—“virtualization” generally relates to a spectrum of definitions rather than to a bright line, and often reflects a design choice with respect to a trade-off between speed and efficiency on the one hand and isolation and universality on the other hand. For example, “full virtualization” is sometimes used to denote a system in which no software components of any form are included in the guest other than those that would be found in a non-virtualized computer; thus, the guest OS 220 could be an off-the-shelf, commercially available OS with no components included specifically to support use in a virtualized environment.
In contrast, another concept, which has yet to achieve a universally accepted definition, is that of “para-virtualization.” As the name implies, a “para-virtualized” system is not “fully” virtualized, but rather the guest is configured in some way to provide certain features that facilitate virtualization. For example, the guest in some para-virtualized systems is designed to avoid hard-to-virtualize operations and configurations, such as by avoiding certain privileged instructions, certain memory address ranges, etc. As another example, many para-virtualized systems include an interface within the guest that enables explicit calls to other components of the virtualization software.
For some, para-virtualization implies that the guest OS 220 (in particular, its kernel) is specifically designed to support such an interface. According to this view, having, for example, an off-the-shelf version of Microsoft Windows XP as the guest OS 220 would not be consistent with the notion of para-virtualization. Others define para-virtualization more broadly to include any guest OS 220 with any code that is specifically intended to provide information directly to any other component of the virtualization software. According to this view, loading a module such as a driver designed to communicate with other virtualization components renders the system para-virtualized, even if the guest OS as such is an off-the-shelf, commercially available OS not specifically designed to support a virtualized computer system. Unless otherwise indicated or apparent, this invention is not restricted to use in systems with any particular “degree” of virtualization and is not to be limited to any particular notion of full or partial (“para-”) virtualization.
In addition to the sometimes fuzzy distinction between full and partial (para-) virtualization, two arrangements of intermediate system-level software layer(s) are in general use—a “hosted” configuration (shown in
As illustrated in
Note that the kernel 600 is not the same as the kernel that will be within the guest OS 220—as is well known, every operating system has its own kernel. Note also that the kernel 600 is part of the “host” platform of the VM/VMM as defined above even though the configuration shown in
In addition to device emulators 370,
The optional console OS in
As a generalization, some form of “virtualization software” executes between system hardware 100 and one or more VMs 200. The virtualization software uses the resources of the system hardware 100, and emulates virtual system hardware 201, on which guest system software 202 and guest applications 260 appear to execute. Thus, virtualization software typically comprises one or more device emulators 370, and either includes or executes in conjunction with some form of system software for accessing and controlling the system hardware 100. The virtualization software may provide full virtualization or partial virtualization. In the non-hosted virtual computer system of
This invention may be used to advantage in both a hosted and/or a non-hosted virtualized computer system, in which the included virtual machine(s) may be fully or para-virtualized, and in which the virtual machine(s) have any number of virtualized processors, which may be of any type (including multi-cored, multi-threaded, or some combination). The invention may also be implemented directly in a computer's primary OS, both where the OS is designed to support virtual machines and where it is not. Moreover, the invention may even be implemented wholly or partially in hardware, for example in processor architectures intended to provide hardware support for virtual machines.
With reference first to
In block 32, at periodic intervals, a check is performed to see if any context in the set is descheduled (e.g., not executing) while any other context in the set is scheduled. There may be a counter associated with each context in a set of contexts, and/or there may be a counter associated with the entire set of contexts. Thus, for example, there may be a counter associated with each virtual CPU (VCPU) and/or with each VM. If any context in the set is descheduled while any other context in the set is scheduled, then (depending on the implementation) the counter for that context and/or the counter for the set of contexts is incremented. If not, the counter(s) may or may not be decremented, again depending on the implementation. That is, if all sibling contexts are scheduled, then the counter for the set of contexts, or any per-context counter that has a counter greater than zero, may either be decremented or left unchanged, depending on how the counter(s) are being implemented.
Block 32 can be implemented as follows. With reference also to
A context that is “idle” or in a formal “halt” state (e.g., as a result of the HLT instruction for the x86 processor architecture) may be considered to be scheduled or running even if it is not. For example, an idle VCPU may execute a guest instruction to halt the VCPU until the next virtual interrupt. Because the guest OS 220 cannot observe the difference between a halted VCPU that is still scheduled and a halted VCPU that has been descheduled transparently by the virtualization platform (e.g., kernel 600 or VMM 300), an idle or halted VCPU can be treated as if it were scheduled or running for purposes of measuring skew. In general, with regard to the discussion of
In block 33 of
System performance can be monitored and each skew threshold value can be adjusted accordingly. Currently, any such adjustment is made manually but automatic adjustments are contemplated. Corrective actions introduced herein include skew reduction policies referred to as “costop” (relaxed and strict), “costart” (relaxed and strict), and “coswap.” Corrective actions are discussed further in conjunction with
With reference next to
In block 42, if skew is detected, an amount of skew is determined. In essence, the amount of time that a context is descheduled while at least one other context in the set is scheduled is measured. In general, for each context in the set, an “event count” is recorded when a context transitions from a scheduled state to a descheduled state, and an event count is recorded when the context transitions from a descheduled state to a scheduled state. The type of event count that is recorded depends on the type of skew metric.
Various skew metrics may be used to indicate the amount of time that one context is ahead of another in terms of execution or scheduling, or conversely to indicate the amount of time that one context is lagging behind another in similar terms.
A skew metric may be a relatively precise measurement of the actual amount of time that a context is descheduled while another context is scheduled, measured in units of real time or processor cycles. For example, for each context (e.g., for each VCPU), the context scheduler may track and record the actual time of each transition between scheduled and descheduled states. Alternatively, the context scheduler may track and record a number of processor cycles at each scheduling transition.
A skew metric may instead be obtained by measuring an attribute that increases, perhaps steadily, with the passage of time. For example, the context scheduler can count the number of memory references, cache misses, and/or instructions executed (e.g., retired) during periods in which a context is descheduled while a sibling context is scheduled. Counts of this nature provide a valuable measure of skew in terms of the amount of work performed (e.g., instructions retired) as opposed to a metric based on elapsed time or processor cycles.
Also, a skew metric may be based on a statistical evaluation of information that is gathered as contexts are scheduled and descheduled. For example, the variance or standard deviation associated with samples of scheduled times provides a useful measure of skew.
Skew may be measured separately for each context in a set of contexts and/or collectively for the entire set of contexts. Thus, for example, skew may be measured for each VCPU and/or for the VM that includes the VCPUs, as well as for other VMs and their respective families of VCPUs.
Skew may be measured per instance or cumulatively. That is, a context's instance skew is associated with a single period during which the context was descheduled while one or more of its siblings were scheduled. The instance skew metric is reset to zero after each such instance.
A context's cumulative skew is accumulated over multiple instances of skew between sibling contexts. In one implementation, a context's cumulative skew is not decreased when the context is scheduled to run along with its siblings. In another implementation, a context's cumulative skew is decreased when the context is scheduled to run along with one of its siblings. In the latter case, a factor may be applied to the amount of the decrease.
For example, with reference to
Furthermore, an aging mechanism of some sort (e.g., an exponentially weighted moving average or the like) can be applied to individual skew values that constitute the cumulative skew value, so that more recent instances of skew are more heavily weighted while past instances of skew are given less weight. Accordingly, a cumulative skew value may “time out” (be reset to zero) if there are no recent instances of skew between sibling contexts. Moreover, a function (e.g., a non-linear function) can be applied to individual skew values that constitute the cumulative skew value, so that some of the individual skew values are weighted differently from the others.
In systems with simultaneous multi-threading (SMT) features (also known as hyper-threading), several contexts can run simultaneously on a single physical processor, while sharing many important execution resources of that processor.
Because the contexts share a resource, they each run at less than 100 percent of full speed. Thus, for example, if one VCPU0 is sharing a hyper-threaded physical processor with another VCPU1 so that each VCPU is running at half speed, while a sibling VCPU2 has sole access to another physical processor and is running at full speed, skew can be introduced because VCPU2 is running faster than the other two VCPUs. In general, the amount of useful work done by a context in a given amount of time may depend on the activity level of the other contexts running on the same physical processor. The present invention can appropriately adjust its measurement of skew to account for SMT effects. For instance, in a skew situation where one context is running at half speed, for example, due to contention on its SMT processor, its skew measure may be advanced at half the full rate. In general, skew per context can be accumulated at less than the full rate to account for effects of SMT when appropriate to do so.
Also, physical processors may run slower due to, for example, power-related throttling (processor power management). For example, physical processors may reduce their clock speed in order to conserve power or reduce heat. As in the SMT example just described, a context running at, for example, half speed due to power throttling may introduce skew between it and its sibling contexts. As in the SMT example, skew per context can be accumulated at less than the full rate to account for effects such as power throttling.
If the condition is not satisfied (e.g., if the skew threshold value is exceeded), a corrective action can be implemented, as described in the figures to follow. The condition may pertain to the instance skew or the cumulative skew discussed above. That is, for example, corrective action may be taken if a context's instance skew exceeds a threshold or if a context's cumulative skew exceeds a threshold value. One condition or threshold value may be defined for instance skew and a different condition or threshold value may be defined for cumulative skew. A different threshold value can also be defined for each context (e.g., per VM and/or per VCPU). System performance can be monitored and each skew threshold value can be adjusted accordingly. Also, as noted in the discussion of
With reference now to
In block 52, in one implementation, all of the coscheduled contexts in the set are descheduled. In another implementation, only a subset of the coscheduled contexts is descheduled. In some instances, only one context (the context with the highest skew metric) may exceed the skew metric applied in block 51, in which case that context would remain scheduled while the other contexts in the set are descheduled. There may be other instances in which multiple contexts exceed the skew metric applied in block 51, in which case only those contexts remain scheduled while the remaining contexts in the set are descheduled. In a sense, the context scheduler forces descheduling of those contexts that are “ahead” while allowing the contexts that are “behind” to continue execution. The former implementation can be referred to as strict costop, while the latter can be referred to as relaxed costop. A choice of relaxed costop versus strict costop can also depend on the particular work being done by the contexts. For example, a strict costop policy may be appropriate for VCPUs that are running critical kernel code, while a relaxed costop policy may be appropriate for VCPUs running userspace code.
In yet another implementation, which may be referred to as “partially relaxed costop,” another threshold value (a relaxed costop threshold), that is less than or equal to the value of the skew threshold applied in block 51, is used to select the members of the subset to be costopped. The costop threshold can be expressed as a fraction or percentage of the skew threshold. In this implementation, in response to block 51, only those coscheduled contexts in the set that have a skew metric that is below the costop threshold are costopped. Thus, if only one of the coscheduled contexts exceeds the skew threshold applied in block 51 and therefore would remain scheduled, other contexts in the set may also remain scheduled even if their respective skew metrics do not exceed the skew threshold. A different costop threshold can also be defined for each context (e.g., per VM and/or per VCPU).
Operation according to a relaxed (or partially relaxed) costop policy can continue according to the relaxed costop policy until all of the skew metrics associated with the set of coscheduled contexts are satisfactory. Costart policies in instances of strict costop are described in conjunction with
Relaxed and partially relaxed costop policies can reduce the number of coscheduling operations relative to strict costop, reducing the number of instances of coscheduling fragmentation and thereby reducing the number of instances in which physical CPUs are not fully utilized. Thus, relaxed costop can improve overall performance. These benefits become more evident as the number of VCPUs per VM increases.
In block 61 of
In block 63 of
Like the skew threshold, the costart threshold can be per instance and/or cumulative. Also, a different costart threshold can also be defined for each context (e.g., per VM and/or per VCPU). Moreover, in a manner similar to that discussed previously herein, a choice of costart policy can also depend on the particular work being done by the contexts—if code, for example, is “coscheduling critical” then a strict costart policy can be implemented, but if the code is “coscheduling irrelevant,” then a relaxed costart policy can be implemented.
In block 64, any context that has a skew metric that is above the costart threshold is scheduled to start before the remaining sibling contexts. For example, a costart threshold of zero requires strict costart, meaning that all of the contexts in the set are to be scheduled and costarted. On the other hand, if the costart threshold is 1.0 (or 100 percent), then only those contexts with a skew metric that is greater than the skew threshold are to be scheduled and costarted. A costart threshold with an intermediate value between zero and 100 percent means that only those contexts with a skew metric that exceeds a corresponding percentage of the skew threshold are to be scheduled and costarted. This latter policy can be referred to as partially relaxed costart.
The costart threshold (expressed as a fraction or percentage of the skew threshold) can be adjusted to improve system performance. Thus, the costart threshold provides explicit control over which of the contexts are to be scheduled and costarted.
When operating under a policy of relaxed or partially relaxed costart, the context scheduler will opportunistically schedule all contexts when sufficient resources are available. That is, in one implementation, the context scheduler is required to costart the subset of contexts that satisfy the costart threshold and may costart the other contexts in the set of sibling contexts.
Relaxed and partially relaxed costart can reduce the number of instances of coscheduling fragmentation because fewer physical CPUs would be needed to continue execution of a VM, for example. Thus, relaxed costop can improve overall performance.
In block 72, in general, each context in the set is rapidly and alternately scheduled and descheduled (started and stopped) on a single physical CPU (e.g., CPU 110 of
For example, with reference to
In summary, methods and systems for coscheduling multiple contexts, including the definition, detection, measurement and management of skew between coscheduled contexts, are disclosed. More refined measures of skew, as well as costop, costart and coswap policies that can be implemented to reduce skew, permit relaxed coscheduling, which in turn reduces fragmentation and improves overall performance.
Although the detailed description is directed to a virtual machine environment, the present invention is not limited to being practiced only on virtual machines. In fact, the present invention is applicable to non-virtual computer systems as well. For example, the present invention can be useful for coscheduling multiple cooperating processes or threads comprising a parallel application within a conventional operating system. Furthermore, one embodiment of the present invention can be in software form stored on a DVD, disk file, or some other type of computer-readable medium.